[TOSA] Add ONNX to TOSA ArgMax conversion pass #45
Conversation
ghost
commented
Jul 15, 2022
ghost
commented
- Add support for direct ArgMax mapping from ONNX to TOSA
… into philippb.tosa_argmax
… into philippb.tosa_argmax
ghost
marked this pull request as ready for review
ghost
force-pushed
the
philippb.refactor_torch_to_tosa
branch
4 times, most recently
from
6397a18
to
c334758
Compare
| return op.emitError( | ||
| "Only floating-point datatype legalization currently supported"); | ||
| } | ||
| static bool isSignedInt(Type type) { |
There was a problem hiding this comment.
Wondering if this should be named isTosaSignedInt - I know it is in a convert-to-tosa file, but I still had to read the contents carefully to understand it was limiting to TOSA supported types. SImilar comment for isFloat.
Also wondering if there is any way to reuse Tosa_SignedInt from mlir/include/mlir/Dialect/Tosa/IR/TosaTypesBase.td directly? (I'm guessing not, but wondered if you knew for certain.)
There was a problem hiding this comment.
Yeah, there isn't. They turn into anonymous constraint functions in the .cpp files where they are used. That's the buf thing that my concepts proposal was to solve.
ghost
force-pushed
the
philippb.refactor_torch_to_tosa
branch
from
49ad859
to
f8bdb7a
Compare
ghost
force-pushed
the
philippb.refactor_torch_to_tosa
branch
from
59e583f
to
aac157c
Compare
ljfitz
changed the base branch from
philippb.refactor_torch_to_tosa
to
ferdinand.bring_upstream_onnxtotosa
ljfitz
changed the base branch from
ferdinand.bring_upstream_onnxtotosa
to
feature/onnx_to_torch
There was a problem hiding this comment.
Could you please re-open this PR on https://github.com/xilinx/onnx-mlir and target feature/onnx-to-tosa?
